Automatic Detection of COVID-19 from Ultrasound Data
NOTE: This is an ultrasound data collection initiative for COVID-19. Please help growing the database or contribute new models.
We develop a computer vision approach for the diagnosis of COVID-19 infections from Point-of-care Ultrasound (POCUS) recordings. Find the arXiv preprint here. This is the first approach to automatically detect COVID-19 from ultrasound. Next to the code for our model and our website on https://pocovidscreen.org, we also make a dataset available. This complements the current data collection initiaves that only focus on CT or X-Ray data. The data includes a total of >100 videos and >30 images of 3 classes (COVID-19, pneumonia, healthy lungs). With frame-based prediction models this easily makes up for a dataset of >1000 images. Please help growing the database!
Motivation:
From the machine learning community, ultrasound has not gained much attention in the context of COVID-19 so far, in contrast to CT and X-Ray scanning. Many voices from the medical community, however, have advocated for a more prominent role of ultrasound in the current pandemic (details below).
Why imaging?
Biomedical imaging has the potential to complement conventional diagnostic procedures for COVID (such as RT-PCR or immuno assays). It can provide a fast assessment and guide downstream diagnostic tests, in particular in triage situations. In the current pandemic, prognostic predictions were already shown to help saving the scarce time of doctors by reducing the time spent to make the right decision (Shan et. al., 2020). Two studies reported that CT imaging can detect COVID-19 at higher sensitivity rate compared to RT-PCR (98% vs 71%, Fang et. al., 2020 and 88% vs. 59% Ai et. al., 2020).
Why ultrasound?
However, ultrasound data was shown to be highly correlated with CT, the gold standard for lung diseases. Instead of CT, ultrasound is non-invasive, cheap, portable (bedside execution), repeatable and available in almost all medical facilities. But even for trained doctors detecting COVID-19 from ultrasound data is challenging and time-consuming. Since their time is scarce, there is an urgent need to simplify, fasten & automatize the detection of COVID-19.
This project is a proof of concept, showing that a CNN is able to learn to distinguish between COVID-19, Pneumonia and healthy patients with an accuracy of 89% and sensitivity for COVID of 96%. This is by no means clinically relevant and a lot of further work needs to be done, e.g. on differentiating COVID from other viral pneumonias.
Evidence for ultrasound
- Peer-reviewed publications from the medical community suggesting to use more ultrasound for COVID-19:
- Ultrasound can evidence the same symptoms as CT: (Point-by-point correspondance of CT and ultrasound findings through COVID-19 disease process)
- NIH launched an initiative to use POCUS for risk stratificaton of COVID-19 patients.
Example lung ultrasound images. (A): A typical COVID-19 infected lung, showing small subpleural consolidation and pleural irregularities. (B): A pneumonia infected lung, with dynamic air bronchograms surrounded by alveolar consolidation. (C) Healthy lung.
The lung is normally aerated with horizontal A-lines.
Contribute!
- You can donate your lung ultrasound recordings directly on our website: https://pocovidscreen.org
- Please help us to find more data! Open an
issue if you
identified a promising data source. Please check here whether the data is
already included. Useful contributions are:
- Publications with ultrasound images/videos
- Images/Videos that are available via Creative Commens license (e.g. CC BY-NC-SA) in the web or on YouTube.
- Possible sources are:
- We are mostly looking for healthy lung recordings (at the moment we have more data for COVID than for healthy lungs)
Learn more about the project
- Read our full arXiv manuscript
- Web Interface (https://pocovidscreen.org)
- Read our blogpost
- DevPost
- Watch this video (3min pitch)
Installation
Ultrasound data
Find all details on the current state of the database in the data folder.
Deep learning model (pocovidnet)
Find all details on how to reproduce our experiments and train your models on ultrasound data in the pocovidnet folder.
Web interface (pocovidscreen)
Find all details on how to get started in the pocovidscreen folder.
Current results
Current results of POCOVID-Net were obtained in a 5-fold CV and show an accuracy of 0.89 (balanced accuracy 0.82) across all 3
classes. For COVID-19, we achieve a sensitivity of 96%.
Detailed performances:
Contact
- If you experience problems with the code, please open an issue.
- If you have questions about the project, please reach out:
jborn@ethz.ch.
Citation
The paper is available here.
If you build upon our work or find it useful, please cite our paper:
@article{born2020pocovid,
title={POCOVID-Net: Automatic Detection of COVID-19 From a New Lung Ultrasound Imaging Dataset (POCUS)},
author={Born, Jannis and Br{\"a}ndle, Gabriel and Cossio, Manuel and Disdier, Marion and Goulet, Julie and Roulin, J{\'e}r{\'e}mie and Wiedemann, Nina},
journal={arXiv preprint arXiv:2004.12084},
year={2020}
}